What you will be able to do
- Recognise the signs that a single prompt is trying to do too much
- Define what 'done' looks like, and the criteria for it, before the first stage starts
- Order the stages of a multi-step business request so each one builds on a checked result
- Judge when decomposition adds nothing and a single well-specified prompt is enough
Key concept
Prompt chaining (task decomposition) — Splitting one complex request into a sequence of smaller prompts. Each prompt handles one stage, and its output becomes the input for the next. Each step is easier for Claude to get right, and you can check the work between steps, at the cost of more round trips.
1.When one prompt is doing too much
Many weak results come from a single request that tries to do everything in one go. One example is: "read these fifteen sets of interview notes, find the themes, decide which matter most and write the findings section of our report." That request has at least four jobs in it: extracting, grouping, judging and drafting. If one response has to do all four, any mistake in an early job is buried inside the finished prose.
The first and lightest fix happens inside a single prompt. Anthropic's prompting guidance says to provide instructions as sequential steps, using numbered lists or bullet points, when the order or completeness of steps matters. Claude can then see which steps it has to complete, and in what order.
Sometimes listing the steps is not enough, and the request needs to become several separate prompts. Anthropic's guidance on prompt chaining gives these signals that a chain is worth it: you have a complex request that needs breaking down into steps, you need iterative refinement, you are doing multi-stage analysis, checking the work partway through adds value, or a single prompt produces inconsistent results. A related observation from Anthropic's engineering team: for complex tasks with several considerations, models generally do better when each consideration gets its own call. Each one then gets focused attention.
| When to use prompt chaining | What it looks like in business work |
|---|---|
| You have a complex request that needs breaking down into steps | Turning a stack of interview notes into a report's findings section |
| You need iterative refinement | Drafting a policy summary, then reviewing it and improving it |
| You're doing multi-stage analysis | Agreeing evaluation criteria first, then scoring applications against them |
| Intermediate validation adds value | Checking an outline before the full document is written |
| A single prompt produces inconsistent results | The same all-in-one request returns a different structure each time |
Checkpoint 1 of 5· Check yourself
A team runs the same all-in-one request ('analyse these survey results, pick the top issues and draft recommendations') three times and gets three differently structured answers. What does Anthropic's guidance suggest this signals?
Inconsistent results from one prompt are named directly as a reason to use prompt chaining. Splitting the work into stages makes each step easier and lets you check it.
“A single prompt produces inconsistent results”Source: claude.com
2.Decompose the problem and define 'done' before you start
Decomposition starts before any prompt is written. Anthropic's AI Fluency course for builders describes delegation as breaking the problem down first, then deciding what AI handles at each step. It also recommends writing acceptance tests before any work begins, because they give you and the AI a shared definition of done. Each statement has to be concrete enough that a stranger could tell whether the finished product passes or fails. The course's good example is: "A patient with a basic smartphone can find the current wait time in under 30 seconds without creating an account." Its bad example is "The system is easy to use."
In everyday business work, defining done often means agreeing the criteria first. Suppose a grants manager has to choose between eleven funding applications. If Claude starts scoring straight away, it has to invent the scoring standard while it scores. If the criteria are agreed in a first stage, every later stage has a fixed standard to apply, and the manager can check the criteria before anything depends on them.
The AI Fluency framework names this skill. Description, the competency of communicating clearly with AI, has three parts. Product Description defines the output. Process Description guides the approach. Performance Description shapes how the AI behaves during the collaboration. Decomposing a request is Process Description in practice: you set out the route as well as the destination.
Checkpoint 2 of 5· Exam question
A legal team needs to compare two lengthy contracts to highlight differences. What is the safest decomposition approach with Claude?
Correct answer: B — Extract the key clauses from each contract first, then compare them point by point.
- A. Incorrect. Asking Claude to output a comparison table with both contracts in a single request risks exceeding the context window limit, leading to truncated input or incomplete comparisons. This approach also makes it harder to manage nuanced differences safely.
- B. Correct. Extracting key clauses first reduces complexity, focuses on relevant parts, and minimizes context overflow. This step-by-step decomposition improves accuracy by breaking the task into manageable, point-by-point comparisons.
- C. Incorrect. Summarizing each contract separately may omit important legal details and precise wording differences necessary for accurate comparison. Comparing only summaries is less reliable than directly comparing extracted clauses.
- D. Incorrect. Loading both lengthy contracts into a single context window to generate a detailed line-by-line diff risks exceeding the token limit, resulting in truncated input or an incomplete diff. This approach is less safe than decomposing the task for thorough comparison.
Checkpoint 3 of 5· Match them up
Match each part of the Description competency to what it covers
Tap a term, then the definition that fits it.
The AI Fluency framework splits Description into three parts. Laying out the stages of a task belongs to Process Description.
“It includes Product Description (defining outputs), Process Description (guiding the approach), and Performance Description (shaping AI behavior during collaboration)”Source: academy.claude.com
3.Ordering the stages so each builds on a checked result
The order of stages is not arbitrary. Each stage should produce something small enough to check, and the next stage should build on it. Anthropic's example of a good chain fits most business writing: write an outline of a document, check that the outline meets certain criteria, and then write the document from the outline. The structure is settled and checked before any time goes into full prose.
The AI Fluency course for educators uses the same order for a slide deck. First ask for an initial slide outline and check whether the flow is logical. Then pick one crucial slide to develop in detail together, and finally combine the results into the full deck. In both cases, broad and cheap-to-change work comes first, and detailed work comes after the structure has been checked.
Chaining works best when the task can be cleanly split into fixed subtasks. It also does not need any special tooling. Chains are often built in code, but you can also run one by hand in a chat: send a prompt, read the response, then send the next prompt with that response as input.
Checkpoint 4 of 5· Put it in order
Put the stages of Anthropic's document-writing chain in order
- 1.Check that the outline meets the agreed criteria
- 2.Write the document based on the outline
- 3.Write an outline of the document
The outline is checked before the full document is written, so any problem with the structure is caught while it is still cheap to fix.
“Writing an outline of a document, checking that the outline meets certain criteria, then writing the document based on the outline.”Source: www.anthropic.com
4.When decomposition is the wrong answer
Decomposition has costs: more prompts, more waiting and more places for things to go wrong. Anthropic's engineering guidance recommends finding the simplest solution possible and adding complexity only when it is needed. For many tasks, getting a single call right is enough.
Current models also do much of the step-by-step work themselves. Anthropic's prompting guidance notes that, with adaptive thinking and subagent orchestration, Claude handles most multistep reasoning internally. It says explicit chaining is still useful when you need to inspect intermediate outputs or enforce a specific pipeline structure. Anthropic also reports teams that built elaborate multi-step setups and then found that improving the prompt for a single agent gave equivalent results.
The practical test is: do you need to see, check or control something between steps? If so, split the work into stages. If you only want the steps done in a particular order, a single prompt with a numbered list of steps is usually enough.
Checkpoint 5 of 5· Check yourself
A manager wants Claude to turn a one-page meeting transcript into a five-bullet action list. Nobody needs to check anything partway through. What approach fits Anthropic's guidance best?
Anthropic recommends starting with the simplest solution and adding complexity only when it is needed. Nothing in this task needs to be inspected between stages.
“we recommend finding the simplest solution possible, and only increasing complexity when needed.”Source: www.anthropic.com
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Any request with more than one step should be broken into a chain of separate prompts.Why is that wrong?
Start with the simplest approach that works. Claude handles most multistep reasoning internally, and explicit chaining earns its cost when you need to inspect intermediate outputs or enforce a specific structure.
Covered in When decomposition is the wrong answer
2.Breaking a task into stages is mainly a way to get the answer faster.Why is that wrong?
Chaining adds latency because it takes several calls. What you gain is accuracy and reliability, because each call has an easier task.
Covered in Ordering the stages so each builds on a checked result
3.With a multi-stage task, you can start with the first piece of work and decide the evaluation standard later.Why is that wrong?
Decide what 'done' means, or which criteria apply, before the work starts. That gives you and the AI a shared, testable standard for every later stage.
Covered in Decompose the problem and define 'done' before you start
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Provide instructions as sequential steps using numbered lists or bullet points when the order or completeness of steps matters.”
↩︎ When one prompt is doing too much“still useful when you need to inspect intermediate outputs or enforce a specific pipeline structure”
↩︎ When decomposition is the wrong answer“With adaptive thinking and subagent orchestration, Claude handles most multistep reasoning internally.”
↩︎ Exam trap 1 - 2.https://claude.com/blog/best-practices-for-prompt-engineeringSecondary source
“A single prompt produces inconsistent results”
↩︎ When one prompt is doing too much“you could manually provide the prompts after receiving responses”
↩︎ Ordering the stages so each builds on a checked result“Chaining breaks down complex tasks into smaller sequential steps with separate prompts.”
↩︎ Key concept“Chaining increases latency (multiple API calls) but often dramatically improves accuracy and reliability for complex tasks.”
↩︎ Exam trap 2 - 3.https://www.anthropic.com/engineering/building-effective-agentsSecondary source
“LLMs generally perform better when each consideration is handled by a separate LLM call”
↩︎ When one prompt is doing too much“This workflow is ideal for situations where the task can be easily and cleanly decomposed into fixed subtasks.”
↩︎ Ordering the stages so each builds on a checked result“The main goal is to trade off latency for higher accuracy, by making each LLM call an easier task.”
↩︎ Prediction“Writing an outline of a document, checking that the outline meets certain criteria, then writing the document based on the outline.”
↩︎ Checkpoint“we recommend finding the simplest solution possible, and only increasing complexity when needed.”
↩︎ Checkpoint - 4.https://academy.claude.com/courses/ai-fluency-for-builders/delegation-the-builder-s-toolkitOfficial docs
“Delegation means decomposing the problem first, then deciding what AI handles at each step.”
↩︎ Decompose the problem and define 'done' before you start“Each must be concrete enough that a stranger could tell you whether the finished product passes or fails.”
↩︎ Decompose the problem and define 'done' before you start“Write acceptance tests before code. They give you and AI a shared definition of done.”
↩︎ Exam trap 3 - 5.
“Description is how you communicate effectively with AI systems.”
↩︎ Decompose the problem and define 'done' before you start“It includes Product Description (defining outputs), Process Description (guiding the approach), and Performance Description (shaping AI behavior during collaboration)”
↩︎ Checkpoint - 6.https://academy.claude.com/courses/ai-fluency-for-educators/applying-ai-fluency-to-learning-materials-and-assignmentsOfficial docs
“Pick one crucial slide to develop in detail together”
↩︎ Ordering the stages so each builds on a checked result“Combine the results to build out the full deck”
↩︎ Ordering the stages so each builds on a checked result - 7.
“improved prompting on a single agent achieved equivalent results”
↩︎ When decomposition is the wrong answer